Electric vehicle charging load prediction method in low data quality scene

Abstract

The invention discloses an electric vehicle charging load prediction method in a low data quality scene, and relates to the technical field of electric vehicles, and the method comprises the steps: obtaining a first feature; setting a dynamic space-time weight factor according to the first feature; according to the dynamic space-time weight factors, vacancy values in the space-time associated charging power matrix are complemented, and complete target charging power matrix data are obtained; constructing a three-dimensional input tensor according to the complete target charging power matrix data; inputting the three-dimensional input tensor into a first model to obtain a charging load predicted value; performing trend verification on the charging load prediction value and historical charging load data to obtain a prediction trend; and if the difference value between the prediction trend and the historical trend exceeds a preset threshold value, correcting the charging load prediction value based on the change rate of the historical trend to obtain a target electric vehicle charging load prediction result. According to the method, high-precision load prediction can be realized, and powerful support is provided for power grid demand response.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electric vehicles, in particular to an electric vehicle charging load prediction method in a low data quality scenario. BACKGROUND

[0002] As a green and environmentally friendly means of transportation, the number of electric vehicles is growing explosively, but their large-scale access to the power grid poses challenges to stable operation and demand response. Accurate prediction of short-term charging load at the hourly level within the next day is crucial for power grid optimization and dispatch, effective demand response strategy development, and ensuring the safe and stable operation of the power grid. However, in practice, charging load data often has gaps due to charging pile failures, communication interruptions, storage problems, and human operations, and poor data quality seriously affects prediction accuracy, thereby hindering the implementation of demand response strategies. Therefore, how to achieve short-term charging load prediction in a low data quality scenario has become an important research topic in the field of electric power systems and electric vehicles.

[0003] Current research on electric vehicle charging load prediction has made certain achievements. Traditional prediction methods mainly include time series analysis methods and machine learning methods (such as support vector machines). In recent years, deep learning methods have been widely used in this field. Long short-term memory networks (LSTM) and convolutional neural networks (CNN) are commonly used to process complex time series data. To further extract long-term dependencies, a new deep learning architecture called Informer is introduced. Its sparse improvement based on self-attention mechanism can effectively capture long-term dependencies and reduce computational complexity. At the same time, through the probabilistic sparse self-attention mechanism, it reduces the processing of redundant information, achieving a good balance between prediction accuracy and time complexity, and is suitable for large-scale and complex charging load data scenarios.

[0004] Traditional time series analysis methods require high data integrity and stationarity, making it difficult to effectively handle data gaps. Traditional machine learning methods can handle data gaps to some extent, but when faced with large-scale and complex charging load data, they lack prediction accuracy and generalization ability. SUMMARY

[0005] The application provides an electric vehicle charging load prediction method in a low data quality scenario, which can achieve high-precision load prediction and provide strong support for power grid demand response.

[0006] To achieve the above purpose, the application adopts the following technical solutions: In a first aspect, the application provides an electric vehicle charging load prediction method in a low data quality scenario, which comprises: obtaining a first feature that affects charging load; setting a dynamic spatio-temporal weight factor according to the first feature that affects charging load; constructing a charging power matrix associated with space-time; completing missing values in the charging power matrix associated with space-time according to the dynamic space-time weight factor to obtain complete target charging power matrix data; constructing a three-dimensional input tensor according to the complete target charging power matrix data; inputting the three-dimensional input tensor into the first model to obtain a charging load prediction value; performing trend verification on the charging load prediction value and historical charging load data to obtain a prediction trend; if a difference between the prediction trend and a historical trend exceeds a preset threshold, performing linear correction on the charging load prediction value based on a change rate of the historical trend to obtain a target electric vehicle charging load prediction result.

[0007] In some possible implementation manners, the obtaining of the first feature affecting the charging load comprises: obtaining first charging data from a plurality of electric vehicle charging stations and charging piles; obtaining multi-source associated influence data; identifying outliers in the first charging data and the multi-source associated influence data by using an outlier detection algorithm, and removing the outliers; performing encoding conversion on text type associated influence data in the first charging data and the multi-source associated influence data after removing the outliers, and unifying a numerical value range to construct a first numerical feature set; constructing a second numerical feature set according to non-text type associated influence data in the first charging data and the multi-source associated influence data after removing the outliers; performing screening on the first numerical feature set and the second numerical feature set by using feature correlation analysis to obtain the first feature affecting the charging load.

[0008] In some possible implementation manners, the completing of the missing values in the charging power matrix associated with space-time according to the dynamic space-time weight factor to obtain the complete target charging power matrix data comprises: constructing a low-rank matrix completion optimization model based on the dynamic space-time weight factor; solving the low-rank matrix completion optimization model by using a singular value threshold algorithm to obtain a target value, replacing the missing values in the charging power matrix associated with space-time with the target value to obtain a completed charging power matrix; verifying the completed charging power matrix to obtain a verification pass result, and taking the completed charging power matrix as the target charging power matrix data.

[0009] In some possible implementation manners, the method further comprises: constructing an initial neural network model; The initial neural network model is optimized by using a flower pollination algorithm to obtain a first model.

[0010] In some possible implementation manners, the method further includes: a full-scenario multi-dimensional evaluation system including an error index, a scenario adaptability index, a robustness index, and an economic index is constructed; weights of the evaluation indexes are determined by using an analytic hierarchy process; actual scores under the evaluation indexes are obtained according to the target electric vehicle charging load prediction result; a comprehensive evaluation result is obtained according to the weights of the evaluation indexes and the actual scores under the evaluation indexes; if the comprehensive evaluation result reaches a preset standard, a demand response strategy is formulated according to the target electric vehicle charging load prediction result.

[0011] In some possible implementation manners, the demand response strategy is formulated according to the target electric vehicle charging load prediction result, including: an early warning level is constructed according to the target electric vehicle charging load prediction result and a maximum bearable load threshold of a power grid, and a division standard of the early warning level is dynamically adjusted according to a proportion of a predicted load in the target electric vehicle charging load prediction result that exceeds the maximum bearable load threshold of the power grid and a load difference change rate; user stratification is performed in combination with a user portrait and a user response behavior prediction model to obtain a user level; the user portrait includes a user charging frequency, a charging time period preference, a charging power demand, a price sensitivity degree, and a historical response record, and the user response behavior prediction model predicts a response probability of each level user under different incentive strategies based on historical data; an incentive intensity is dynamically adjusted according to the response probability of the user, and a demand response strategy is formulated according to the incentive intensity, the early warning level, and the user level.

[0012] In some possible implementation manners, the dynamic spatiotemporal weight factor includes a region weight and a time period weight; the region weight is dynamically adjusted according to a historical load contribution degree and an operation state of a charging station in a corresponding region in the first feature of the charging load impact, and the time period weight is dynamically configured according to a charging load fluctuation feature in different time periods and a real-time electricity demand change in the first feature of the charging load impact.

[0013] In a second aspect, the application provides a device for predicting electric vehicle charging load in a low data quality scenario, the device including: an acquisition module configured to acquire a first feature of charging load impact; The supplement module is configured to set a dynamic space-time weight factor according to the first characteristic of the charging load influence, construct a space-time correlated charging power matrix, and supplement the missing values in the space-time correlated charging power matrix according to the dynamic space-time weight factor to obtain complete target charging power matrix data. The prediction module is configured to construct a three-dimensional input tensor according to the complete target charging power matrix data, input the three-dimensional input tensor into the first model to obtain a charging load prediction value, perform trend verification on the charging load prediction value and historical charging load data to obtain a prediction trend, and perform linear correction on the charging load prediction value based on a change rate of the historical trend to obtain a target electric vehicle charging load prediction result if a difference between the prediction trend and the historical trend exceeds a preset threshold.

[0014] In a third aspect, the present application provides a computing device including a memory and a processor. The memory stores one or more computer programs including instructions, and the instructions, when executed by the processor, cause the computing device to perform the method of any one of the first aspect.

[0015] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program for performing the method of any one of the first aspect.

[0016] In a fifth aspect, the present application provides a computer program product including one or more computer instructions, which, when executed by a computer, perform the method of any one of the first aspect.

[0017] From the above technical solutions, the present application has at least the following beneficial effects: In the present application, the first characteristic of the charging load influence is obtained, a dynamic space-time weight factor is set according to the first characteristic of the charging load influence, a space-time correlated charging power matrix is constructed, missing values in the space-time correlated charging power matrix are supplemented according to the dynamic space-time weight factor to obtain complete target charging power matrix data, a three-dimensional input tensor is constructed according to the complete target charging power matrix data, the three-dimensional input tensor is input into the first model to obtain a charging load prediction value, trend verification is performed on the charging load prediction value and historical charging load data to obtain a prediction trend, and linear correction is performed on the charging load prediction value based on a change rate of the historical trend to obtain a target electric vehicle charging load prediction result if a difference between the prediction trend and the historical trend exceeds a preset threshold.

[0018] In the prior art, simple interpolation methods or single models are often used to directly predict the charging load, without fully considering the spatiotemporal dynamic change characteristics of the charging load related features, lacking of pertinence in filling the missing values of the charging power matrix, and lacking of checking and correcting links for the predicted results and historical trends, resulting in limited prediction result accuracy, difficulty in reflecting the actual charging load change law, and bringing adverse effects to subsequent power grid demand response work based on the prediction results. It can be seen that, by obtaining the first feature affecting the charging load and setting a dynamic spatiotemporal weight factor, the spatiotemporal change law of the charging load can be captured faster, the missing values of the charging power matrix are filled, and a three-dimensional input tensor more in line with the actual situation is constructed; in combination with the first model prediction and the checking and correction of the predicted trend and the historical trend, the accuracy and reliability of the charging load prediction results are guaranteed in multiple dimensions, providing strong data support for efficient demand response and optimal resource allocation of the power grid.

[0019] It should be understood that the description of technical features, technical solutions, beneficial effects or similar language in this application does not imply that all features and advantages can be achieved in any single embodiment. On the contrary, it can be understood that the description of a feature or beneficial effect means that the specific technical feature, technical solution or beneficial effect is included in at least one embodiment. Therefore, the description of technical features, technical solutions or beneficial effects in this specification does not necessarily refer to the same embodiment. Further, the technical features, technical solutions and beneficial effects described in this embodiment can be combined in any appropriate manner. Those skilled in the art will understand that the embodiments can be implemented without one or more specific technical features, technical solutions or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects can be identified in specific embodiments that do not embody all embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of a low-data-quality scenario electric vehicle charging load prediction method provided by an embodiment of the application; Figure 2 A schematic diagram of a prediction result of an electric vehicle charging load provided by an embodiment of the application; Figure 3 A schematic diagram of a prediction error result provided by an embodiment of the application; Figure 4 A schematic diagram of a low-data-quality scenario electric vehicle charging load prediction device provided by an embodiment of the application; Figure 5 A schematic diagram of a computing device provided by an embodiment of the application. DETAILED DESCRIPTION

[0021] The terms "first", "second", and "third" and the like in the specification and the drawings of the present application are used to distinguish different objects, and are not used to define a particular order.

[0022] In the embodiments of the present application, the words such as "exemplary" or "for example" are used to mean serving as an example, instance, or illustration, any embodiment or aspect described as "exemplary" or "for example" in the embodiments of the present application is not necessarily to be construed as preferred or advantageous over other embodiments or aspects. Rather, use of the words such as "exemplary" or "for example" is intended to present concepts in a concrete manner.

[0023] For the sake of clear and concise description of each embodiment below, a brief introduction of the related art is given first: As green and environmentally friendly transportation tools, the number of electric vehicles is growing explosively. However, the large-scale access of electric vehicles also brings new challenges to the power grid, especially the uncertainty of charging load, which puts higher requirements on the stable operation and demand response of the power grid. Accurate prediction of short-term charging load of electric vehicles, especially the hourly charging load distribution in the future, is the key to power grid optimization and demand response strategy, and is one of the factors to ensure the safe and stable operation of the power grid.

[0024] In actual application scenarios, electric vehicle charging load data often faces quality problems, and data vacancy is a typical case. The reasons for data vacancy are various: it may be charging pile failure, it may be communication interruption, such as communication link failure between charging pile and data server, resulting in part of the data unable to upload, it may be data storage problem, and it may also be affected by human factors. Such data quality problems will seriously reduce the accuracy of electric vehicle charging load prediction, and further hinder the effective formulation and landing of power grid demand response strategy. Therefore, how to accurately predict the short-term charging load of electric vehicles under the condition of poor data quality has become an important research direction in the field of electric power system and electric vehicles.

[0025] At present, the research on electric vehicle charging load prediction has made certain progress. Traditional prediction methods are mainly divided into time series analysis method and machine learning method (such as support vector machine), but these methods have obvious limitations in charging load prediction under low data quality scenarios: time series analysis method has high requirements for data integrity and stationarity, and it is difficult to effectively handle data vacancy problem; traditional machine learning method can deal with data vacancy to some extent, but when facing large-scale and complex electric vehicle charging load data, its prediction accuracy and generalization ability still have room for improvement.

[0026] In recent years, deep learning methods have been widely applied in this field. Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN) and other methods are often used to process complex time series data. However, these methods cannot effectively extract the long-term dependence of load data. To solve this problem, a new deep learning architecture, Informer, is introduced. Based on the sparsity improvement of self-attention mechanism, Informer can effectively capture long-term dependencies and significantly reduce computational complexity. In addition, Informer reduces the need for redundant information processing through probabilistic sparse self-attention mechanism, further improving model efficiency and performance. Compared with traditional deep learning methods, Informer achieves a better balance between prediction accuracy and time complexity, which is very suitable for large-scale and complex electric vehicle charging load data scenarios. However, the setting of Informer's hyperparameters relies heavily on human experience, which can easily lead to unstable prediction performance.

[0027] Therefore, the embodiment of the present application provides a method for predicting electric vehicle charging load in a low data quality scenario. First, the first feature that affects the charging load is obtained. Then, a dynamic spatio-temporal weight factor is set according to the first feature that affects the charging load. A spatio-temporal correlated charging power matrix is constructed. The missing values in the spatio-temporal correlated charging power matrix are completed according to the dynamic spatio-temporal weight factor to obtain complete target charging power matrix data. A three-dimensional input tensor is constructed according to the complete target charging power matrix data. The three-dimensional input tensor is input into the first model to obtain the charging load prediction value. The charging load prediction value is trend checked with the historical charging load data to obtain the prediction trend. If the difference between the prediction trend and the historical trend exceeds the preset threshold, the charging load prediction value is linearly corrected based on the change rate of the historical trend to obtain the target electric vehicle charging load prediction result. By obtaining the first feature that affects the charging load and setting the dynamic spatio-temporal weight factor, the spatio-temporal variation law of the charging load can be more accurately captured, and the missing values of the charging power matrix can be completed to construct a three-dimensional input tensor that is more in line with reality. Combined with the first model prediction and the checking and correction of the prediction trend and the historical trend, the accuracy and reliability of the charging load prediction result are guaranteed from multiple dimensions, providing strong data support for the efficient development of demand response and optimization of resource allocation of the power grid.

[0028] In order to make the technical solutions of the present application clearer and easier to understand, the following will introduce a method for predicting electric vehicle charging load in a low data quality scenario provided by the embodiment of the present application in conjunction with the drawings. As shown in the figure, it is a flowchart of a method for predicting electric vehicle charging load in a low data quality scenario provided by the embodiment of the present application. Figure 1

[0029] ​In the embodiment of the present application, the method is applied to a processing device. The low-data-quality electric vehicle charging load prediction method comprises the following steps: S101, a first feature affecting the charging load is acquired.

[0030] Firstly, multi-source data collection is carried out. Real-time first charging data is acquired from a plurality of electric vehicle charging stations and charging piles by means of intelligent sensing devices and communication modules built in the charging piles. The first charging data includes charging power, charging start time, charging end time, charging power, and other key information, which can reflect the electric energy parameters in the electric vehicle charging process. Then, multi-source associated influence data is acquired by connecting a meteorological department data platform, a holiday management system, and a local power department electricity price policy release channel. The multi-source associated influence data includes weather conditions (such as temperature, humidity, and light intensity, different weather conditions will affect user travel and charging demand), holiday information (user travel mode changes during holidays, and charging behavior will also change accordingly), and local electricity price policy (time-of-use electricity price policy will guide users to adjust the charging time period) and other data.

[0031] Then, data preprocessing is performed. An outlier detection algorithm is used to identify outliers in the first charging data and the multi-source associated influence data, and the outliers are removed. Specifically, an outlier detection algorithm, such as a 3σ principle detection algorithm based on statistics or an isolated forest algorithm, is used to identify outliers in the first charging data and the multi-source associated influence data. The cause of the outliers may be equipment failure (such as a malfunctioning charging pile sensor causing abnormal charging power data), communication interference (errors in the data transmission process), etc. The identified outliers are removed to ensure the reliability of subsequent data.

[0032] The text type associated influence data in the first charging data and the multi-source associated influence data after removing the outliers is converted and encoded, and the numerical range is unified to construct a first numerical feature set. For example, holiday information and other texts are converted and encoded by using label encoding or one-hot encoding to convert the text information into a numerical form recognizable by a computer. At the same time, all data are normalized or standardized to unify the numerical range, for example, the data are mapped to the interval [0, 1], thereby constructing the first numerical feature set.

[0033] As for the non-text type associated influence data in the first charging data and the multi-source associated influence data after removing the outliers, such as charging power and temperature, which are numerical data, they are directly sorted to construct a second numerical feature set.

[0034] Finally, the feature screening work is carried out. Through the feature correlation analysis method such as Pearson correlation coefficient analysis or mutual information method, the first numerical feature set and the second numerical feature set are screened. The correlation degree between each feature and the charging load is calculated, the features with high correlation with the charging load are retained, and the redundant features with low correlation and small contribution to the charging load prediction are removed, so as to obtain the first features affecting the charging load. Wherein, the high correlation refers to that the absolute value of the Pearson correlation coefficient is greater than or equal to the correlation threshold, and the low correlation refers to that the absolute value of the Pearson correlation coefficient is less than the correlation threshold, and in the present application, the correlation threshold can be set to 0.5; if the absolute value of the Pearson correlation coefficient between a certain feature and the charging load is less than 0.5 (i.e. low correlation), it means that the linear correlation between the feature and the charging load is weak, and the feature cannot provide sufficient effective information reflecting the change rule of the charging load for the prediction model, and the positive help for the prediction of the charging load is small, which belongs to the case of small contribution; even if the absolute value of the Pearson correlation coefficient between a certain feature and the charging load is not less than 0.5 (the correlation is not extremely low), but if it cannot provide additional effective information for the prediction model, such as the time period ID (a simple serial number for the time period), the association between the feature and the charging load only comes from the mapping of the time period, and the independent contribution of the feature to the prediction is extremely low, which also belongs to the case of small contribution. These screened features will be faster used for subsequent construction of the charging load prediction model.

[0035] S102, setting a dynamic space-time weight factor according to the first features affecting the charging load.

[0036] The dynamic space-time weight factor includes a region weight and a time period weight.

[0037] For the region weight, the historical load data of each regional charging station is extracted from the first features affecting the charging load, and the historical load contribution degree is calculated, that is, the proportion of the charging load of the regional charging station in the overall load in the past period of time. And combined with the operation state of each regional charging station, such as the number of available charging piles, the health degree of the equipment (judged by equipment failure records, maintenance period, etc.), and other factors.

[0038] If the historical load contribution degree of a certain regional charging station is high and the operation state is good, it can stably provide charging service, so the region weight is increased; on the contrary, if the historical contribution degree is low or the operation state is poor (such as frequent charging pile failures), the region weight is reduced.

[0039] In this way, the region weight is dynamically adjusted, and the influence degree of different regions of the charging station on the overall charging load can be reflected. For the time period weight, according to the charging load fluctuation characteristics of different time periods in the first feature of the influence on the charging load, the change law of the charging load of each time period (such as early peak, midday, late peak, night, etc.) in history is analyzed, such as the charging load is usually high and fluctuates greatly in the late peak period. Combined with the real-time power demand change, the time period weight is dynamically configured by monitoring the power load of the power grid and the real-time data of the user charging request. In the time period of high charging load and tight real-time power demand, the time period weight is increased to highlight the influence of the charging load on the whole; in the time period of low charging load and loose real-time power demand, the time period weight is reduced.

[0040] By setting the dynamic space-time weight factor, the importance of different regions and different time periods in subsequent analysis and prediction can be dynamically adjusted according to the actual influencing factors of the charging load, so that the subsequent processing based on these weights is more in line with the actual charging load change.

[0041] Compared with the fixed weight method, dynamically adjusting the region weight and the time period weight can more flexibly and accurately reflect the real influence of different regions and time periods on the charging load, improve the description accuracy of the influencing factors of the charging load, and provide more suitable weight basis for the subsequent steps, thereby improving the accuracy of the subsequent charging load related processing.

[0042] S103, constructing a charging power matrix associated with space and time.

[0043] Firstly, the original charging power data of each region charging station in different time periods is collected. Then, the dynamic space-time weight factor set in S102 is used to perform weighted processing on the original data. For the charging power data of each region and each time period, the corresponding region weight and time period weight are multiplied to strengthen the effect of the charging power data of the region and time period that has a large influence on the charging load.

[0044] After that, the charging power data of each region and each time period after the weighted processing is integrated to construct a two-dimensional charging power matrix according to the dimensions of the region and the time period. Among them, the rows of the matrix can represent different regions, the columns can represent different time periods, and each element in the matrix can represent the charging power value of the corresponding region in the corresponding time period after the space-time weight adjustment. The space-time associated charging power matrix constructed in this way can simultaneously reflect the space (region) and time (time period) correlation characteristics of the charging load.

[0045] The constructed space-time correlated charging power matrix integrates the space-time correlation characteristics of the charging load in one matrix, provides a basic data structure containing space-time correlation information for subsequent steps of charging power matrix missing value completion and charging load prediction, and facilitates subsequent models to learn and mine the space-time law of the charging load.

[0046] The constructed space-time correlated charging power matrix integrates the information in the space-time dimension, so that the correlation in space and time can be considered simultaneously when subsequent charging power data is processed, which helps to more comprehensively grasp the change law of the charging load, lays a good data foundation for subsequent charging load prediction tasks, and thus improves the accuracy and reliability of the charging load prediction.

[0047] S104, according to the dynamic space-time weight factor, the missing values in the space-time correlated charging power matrix are completed to obtain complete target charging power matrix data.

[0048] A low-rank matrix completion optimization model based on the dynamic space-time weight factor is constructed, a singular value threshold algorithm is used to solve the low-rank matrix completion optimization model to obtain a target value, the target value is replaced in the missing values in the space-time correlated charging power matrix to obtain a completed charging power matrix. The completed charging power matrix is verified to obtain a result of passing the verification, and the completed charging power matrix is taken as the target charging power matrix data. The specific steps and related formulas are as follows: First, a low-rank matrix completion optimization model based on a dynamic space-time weight factor is constructed. Let the space-time correlated charging power matrix be (M is the space-time correlated charging power matrix, m is the number of regions, n is the number of time periods, R is a mathematical standard symbol, and is used to indicate that the elements in the matrix or vector discussed are real numbers), wherein there are missing values in the charging power matrix, and the index set of the known elements in the space-time correlated charging power matrix is denoted as denotes the index set of the missing elements in the space-time correlated charging power matrix.

[0049] In the dynamic space-time weight factor, the full-region weight is , and the full-time period weight is , so the low-rank matrix completion optimization model integrating the dynamic space-time weight is:

[0050]

[0051] wherein, denotes finding a matrix X such that the rank of the matrix X is as small as possible, is a low-rank matrix to be completed, ​Let X be the element in the i-th row and j-th column, which is the element of the matrix to be completed at that position. charging power matrix with spatiotemporal correlation The known element in the i-th row and j-th column, It represents the rank of the matrix, reflecting the low-rank constraint; To balance the low-rank constraints and the regularization parameters of the data fitting term; This is the region weight corresponding to region i, where i is an integer, and , This is the time period weight corresponding to time period j, where j is an integer, and , and The integration of dynamic spatiotemporal weights makes the fitting of known elements more focused on regions and time periods with high weights, where T is the transpose.

[0052] Next, the singular value thresholding algorithm is used to solve the low-rank matrix completion optimization model. For the matrix... The singular value decomposition is performed using the following formula:

[0053]

[0054] in, It is the first orthogonal matrix. It is the second orthogonal matrix. It is the transpose of the second orthogonal matrix. Let X be a diagonal matrix, and let X be the singular values ​​of the low-rank matrix X to be completed, with these singular values ​​arranged in descending order. For a detailed description of a diagonal matrix, , , ···, Let X be the singular values ​​of X, and k be the effective rank-related dimension of the matrix, i.e., the number of non-zero singular values ​​of X. It is closely related to the rank of the matrix and reflects the effective information dimension of the matrix. .

[0055] The singular value thresholding algorithm achieves low-rank approximation by thresholding singular values, that is:

[0056] in, diagonal matrix The diagonal matrix obtained after thresholding The threshold parameter is determined through iterative optimization.

[0057] Solving for the completed matrix yields the solution. The completed matrix Chinese correspondence The element value of the index, i.e., the target value, is calculated by the formula:

[0058] The corresponding element value of the index is replaced with the missing value in the space-time associated charging power matrix M to obtain a preliminary completed charging power matrix M', and the calculation formula of the element in the charging power matrix M' is:

[0059] Finally, the completed charging power matrix is checked. Select some known data (non-missing data) in the completed matrix to calculate the mean square error , and the formula is as follows:

[0060] If the MSE is less than a preset threshold , it is considered that the completion effect meets the requirements; otherwise, the model parameters (such as , , etc.) are adjusted to solve again. After checking, the complete target charging power matrix data is obtained.

[0061] By constructing a low-rank matrix completion optimization model incorporating dynamic space-time weight factors, and using the singular value threshold algorithm for solving and combining with mean square error checking, the missing values of the space-time associated charging power matrix are completed, providing complete and high-quality charging power data support for subsequent charging load prediction and other links.

[0062] S105, constructing a three-dimensional input tensor according to the complete target charging power matrix data.

[0063] First of all, it is clear that the complete target charging power matrix is denoted as , which is a matrix used to describe the charging power situation in different areas and different time periods, which includes the charging power information of m areas (m is the number of areas) and n time periods (n is the number of time periods). In order to make this matrix be able to be processed by the subsequent first model, it needs to be constructed into a three-dimensional input tensor. The first step is to extend the dimension of .

[0064] Because in the charging load prediction, not only the charging power of the current area and the current time period will have an impact, but also the charging power of the historical time will have an effect on the prediction result due to the time sequence association. Therefore, the application selects the charging power data of T1 consecutive time windows, such as selecting the charging power data of the past T1 hours with an hour as the time window.

[0065] Then, for each region i (i takes values from 1 to m) and each time period j (j takes values from 1 to n) , collect their charging power data in these consecutive T1 time windows, denoted as (where t takes values from 1 to T1) .

[0066] Finally, based on these collected data, construct a three-dimensional input tensor . This three-dimensional input tensor is a tensor with m region dimensions, n time period dimensions, and T1 time window dimensions. Each element in the three-dimensional input tensor corresponds to the charging power data of region i, time period j, and the t1th time window . The three-dimensional input tensor constructed in this way can simultaneously reflect the information of charging load in the region, time period, and time sequence dimensions, providing the necessary data basis for subsequent models to capture the spatiotemporal and time sequence correlation of charging load.

[0067] S106, input the three-dimensional input tensor into the first model to obtain the charging load prediction value.

[0068] Construct an initial neural network model, optimize the initial neural network model using the flower pollination algorithm to obtain the first model, specifically including: The initial neural network model adopts a network architecture based on attention mechanism, which is the Informer model. The core is to capture the long-term dependence of charging load data through self-attention mechanism.

[0069] The overall architecture of the Informer model consists of an encoder and a decoder. The model input covers charging data such as charging power, charging start time, charging end time, and charging capacity, as well as external influencing factor data related to electric vehicle load, such as weather conditions (including temperature, humidity, wind speed, etc.), holiday arrangement information, and local electricity price policy, etc.

[0070] Both the encoder and the decoder are composed of multiple neural network modules, which include multi-head probabilistic sparse self-attention layers and convolution distillation layers. The model first performs time encoding on the input time series data, and then uses multiple neural network modules composed of multi-head probabilistic sparse self-attention layers and convolution layers to perform deep feature extraction. In the decoder stage, the variables related to electric vehicle charging load are input into the model, which are operated with the deep feature maps extracted by the encoder to capture the correlation between the features. Finally, the prediction result of the electric vehicle charging load value is obtained through the fully connected layer.

[0071] In the multi-head probability sparse self-attention layer, redundant calculations are reduced by using a query matrix (Q), so that the time complexity is reduced from O(L 2 ) of the traditional self-attention to O(LlnL). The formula is as follows:

[0072] wherein, is the filtered query matrix, K is the key matrix, is the transpose of the key matrix, V is the value matrix, and L represents the sequence length, is the dimension of the key matrix K, is a function for normalizing The structure of the initial neural network model includes an encoder and a decoder part. The encoder extracts features from the input three-dimensional tensor, and the decoder generates the charge load prediction result based on the extracted features.

[0073] In the overall architecture of the Informer, the encoder uses sparse self-attention instead of the traditional self-attention mechanism. This method can reduce the size of the network and increase the robustness of the model by layer-by-layer stacking. The decoder receives a long sequence input, fills the target element with zero, and combines the feature maps through weighted attention to quickly output the element in a generative manner.

[0074] The main function of the encoder is to extract the long-distance correlation and deep features of the input sequence. In the process of probability sparse self-attention calculation, the query vector with low attention value is directly averaged. In order to make the model structure pay more attention to the dominant attention in the deep feature extraction process and highlight its role, we use convolution and pooling operations to distill the self-attention feature map. In this way, the dominant features with dominant role can be extracted with privilege. The feature transmission process from the rth layer to the r+1th layer is as follows:

[0075] In the formula, is the output feature of the r+1th layer at the tth moment, is the input feature of the rth layer at the tth moment, Conv is a one-dimensional convolution operation used to extract local features; is an exponential linear unit; MaxPool is a maximum pooling operation; is an attention block.

[0076] In the decoder part, a standard decoder structure is adopted. It consists of two identical multi-head attention layers. This structure is used to alleviate the speed drop problem in long sequence generation inference process. First, the hidden multi-head attention layer (Masked Multi-Head Attention) is used to avoid the leakage of future information in autoregressive prediction. Then, the deep features output by the encoder are spliced with the input features of the decoder. The splicing formula is as follows:

[0077] In the formula: is the spliced feature input into the decoder at the t-th moment, is the token-related feature at the t-th moment (token can be understood as a unit for identifying or conveying specific information), is the input feature at the t-th moment, excluding the spliced feature and the token-related feature, is the length parameter related to the token, used to represent the sequence length of the token-related feature, is the length parameter of the token-related feature (such as the target-related feature), is a real number space with a dimension of , that is, the space dimension of the spliced feature vector, and Concat is the splicing operation that splices the features of the encoder and the decoder together; is the dimension of the model, representing the dimension of the feature vector.

[0078] In the probability sparse self-attention, the hidden multi-head attention mechanism (masked multi-head attention) is adopted. This mechanism avoids the autoregressive problem of paying attention to the next position at each position. Finally, the preliminary charging load prediction result is output through the full connection layer.

[0079] Secondly, in order to improve the prediction accuracy and generalization ability of the Informer deep neural network in short-term electric vehicle charging load prediction, the Flower Pollination Algorithm (FPA) is introduced to optimize the key hyperparameters of the model. First, the hyperparameters of the Informer model (including the number of multi-head attention heads, the number of encoder layers, the number of decoder layers, and the learning rate) are encoded into individual position vectors of the Flower Pollination Algorithm. Each vector corresponds to a set of parameter combinations to be evaluated. Then, the population is randomly initialized to be uniformly distributed in the parameter search space. During optimization, the algorithm performs global pollination or local pollination with a certain probability p (in this application, p=0.8 can be taken). In the global pollination stage, long-distance jump steps are generated using Levy flight as follows:

[0080] wherein: is the position of the s th individual in the d+1 th generation, is the position of the s th individual in the d th generation, is the step control factor; L( ) represents a random step subject to a Levy distribution; is the optimal solution of the current population; in the local pollination stage, two individuals and are randomly selected to perform short-distance information exchange and update, and the update formula is:

[0081] wherein: is the position of the f th individual randomly selected in the d th generation, is the position of the z th individual randomly selected in the d th generation, and ε is a random number subject to a uniform distribution.

[0082] After each parameter update, the corresponding Informer model is trained on the training set, and the prediction error (MAPE) is calculated on the validation set, which is used as the fitness value; if the fitness is better than the historical optimum of the current individual, the replacement is performed. The algorithm is terminated when the maximum number of iterations or the fitness convergence condition is met, and the global optimal parameter combination is output, which is used for training and testing of the final Informer model. Through this method, the balance between global optimality and convergence speed of parameter search can be achieved, thereby improving the prediction accuracy and stability of the model for electric vehicle charging load.

[0083] The initial neural network model is optimized by the flower pollination algorithm to obtain an optimized first model. The three-dimensional input tensor constructed in S105 is input into the optimized first model. The encoder first extracts the spatio-temporal features of the three-dimensional input tensor, captures the spatial correlation of regional charging load and the temporal dependence of different time periods through multi-head probabilistic sparse self-attention, and then compresses the feature dimension through the convolution distillation layer. The decoder receives the vector obtained by splicing the deep features output by the encoder and the multi-source correlation features, matches the feature correlation through the hidden multi-head attention layer, and finally outputs the hourly charging load prediction value for the next 24 hours through the full connection layer to complete the prediction calculation.

[0084] S107, trend verification is performed on the charging load prediction value and the historical charging load data to obtain a prediction trend.

[0085] First, sufficient historical charging load data is collected. These data need to cover charging load conditions in different time spans, different seasons and different special periods, ensuring that the data are comprehensive and representative.

[0086] Then, the trend of the predicted charging load and the trend of the historical charging load data are extracted respectively. For the historical charging load data, a time series analysis method such as a moving average method is used to calculate the average charging load in a certain time window, and then the long-term trend and the short-term fluctuation trend of the historical charging load are obtained. For the predicted charging load, the same time granularity and analysis method as the historical data are used to extract the predicted load trend.

[0087] Then, trend verification is performed. The predicted trend is compared with the historical trend in multiple dimensions, including the direction of the trend, the change amplitude of the trend, and the periodic characteristics of the trend, etc. The difference between the two can be observed intuitively by drawing a trend comparison chart, or the matching degree of the predicted trend and the historical trend can be quantified by calculating a trend similarity index such as a cosine similarity. Through these verification operations, the fit of the predicted trend and the historical trend is determined, and a detailed description of the predicted trend is obtained to provide a basis for the subsequent steps.

[0088] S108, if the difference between the predicted trend and the historical trend exceeds the preset threshold, linearly correcting the predicted charging load value based on the change rate of the historical trend to obtain a target electric vehicle charging load prediction result.

[0089] First, a preset threshold is determined. The setting of this threshold needs to consider factors such as the fluctuation range of the historical charging load data and the accuracy requirement of the prediction task. The difference distribution between the historical trend and the ideal trend (or multiple reliable predicted trends) can be calculated by statistical analysis of the historical data, and an appropriate quantile (such as the 90% quantile) is selected as the preset threshold to ensure that when the difference between the predicted trend and the historical trend exceeds the threshold, it means that the predicted trend has a large deviation and needs to be corrected.

[0090] When the difference between the predicted trend and the historical trend exceeds the preset threshold, the change rate of the historical trend is calculated.

[0091] The change rate of the historical trend can be obtained by linear fitting of the historical charging load data. Let the fitting straight line of the historical charging load data in the time interval be y=k1x+b, where the slope k1 is the change rate of the historical trend, which reflects the speed and direction of the change of the charging load over time. Then, the predicted charging load value is linearly corrected based on the change rate of the historical trend. Let the predicted charging load value be , and the corresponding time is t. According to the change rate k1 of the historical trend, the corrected charging load prediction value is calculated as , and the correction formula can be expressed as (where is the time difference from the historical reference time to the prediction time, which can be adjusted according to the actual trend changes). Such linear correction can utilize the change law of the historical trend to reasonably correct the prediction value with large deviation, so as to make it more consistent with the actual load change trend. Finally, the target electric vehicle charging load prediction result after linear correction is output. At the same time, the reasons for correction (i.e. the difference between the prediction trend and the historical trend exceeds the threshold) and the historical trend change rate used in the correction process are recorded, so as to evaluate the correction effect and optimize the model in the future.

[0092] In some embodiments, in order to comprehensively evaluate the electric vehicle charging load prediction result, a full-scenario multi-dimensional evaluation system including error indicators, scenario adaptability indicators, robustness indicators and economic indicators is constructed; the weights of each evaluation indicator are determined by the analytic hierarchy process; the actual scores under each evaluation layer indicator are obtained according to the target electric vehicle charging load prediction result; the comprehensive evaluation result is obtained according to the weights of each evaluation indicator and the actual scores under each evaluation layer indicator; if the comprehensive evaluation result reaches the preset standard, the demand response strategy is formulated according to the target electric vehicle charging load prediction result.

[0093] The evaluation result is usually a comprehensive score, which is obtained by weighting calculation through the actual scores of each evaluation indicator combined with the weights determined by the analytic hierarchy process. For example, the comprehensive score can be a 0-100 numerical value, and the higher the score, the better the performance of the charging load prediction result in the full-scenario multi-dimensional. The preset standard is an artificially set threshold value for judging whether the prediction result meets the requirements. For example, it is set that the comprehensive score of 80 points and above reaches the preset standard. Specific example Assuming that: the error indicator weight is 0.3, the actual score is 90 points (indicating that the prediction error is small, such as the average absolute error is much lower than the historical average error); the scenario adaptability indicator weight is 0.2, the actual score is 85 points (indicating that the prediction result can be well adapted in different scenarios, such as weekdays and holidays); the robustness indicator weight is 0.3, the actual score is 75 points (meaning that when the input data has small fluctuations, the prediction result will not have large deviation); the economic indicator weight is 0.2, the actual score is 95 points (indicating that the charging price strategy based on the prediction result can bring good economic benefits). According to the comprehensive evaluation calculation method: (90x0.3+85x0.2+75x0.3+95x0.2=27+17+22.5+19=85.5) points. Because 85.5 points ≥ 80 points, it reaches the preset standard, so the demand response strategy can be formulated according to the electric vehicle charging load prediction result.

[0094] For the error index, it is the key to measure the deviation degree between the predicted value and the true value. The error index includes mean absolute error (MAE), mean absolute percentage error (MAPE), root mean squared error (RMSE) and prediction trend accuracy rate. The scene adaptability index includes prediction performance under different data missing rates and different weather scenes. The robustness index includes prediction stability under data noise interference and hardware failure simulation scenes. The economic index includes changes in power grid scheduling cost and user charging cost after formulating demand response strategies based on prediction results.

[0095] The MAE, MAPE and RMSE calculation formulas are as follows: The mean absolute error (MAE) can intuitively reflect the average size of the prediction error, and the calculation formula is:

[0096] The mean absolute percentage error Then it focuses on reflecting the proportion of error in the true value, and can more intuitively reflect the relative error situation, and the formula is:

[0097] The root mean squared error RMSE is more sensitive to large errors, and can highlight the influence of outliers on prediction results, and the formula is:

[0098] In the formula, C is the number of prediction points, c is greater than or equal to 1 and less than or equal to C, is the predicted value of the cth electric vehicle charging load, is the true value of the cth electric vehicle charging load, MAE is the mean absolute error, MAPE is the mean absolute percentage error, and RMSE is the root mean squared error.

[0099] In addition, there is a prediction trend accuracy rate, which is used to judge the degree of agreement between the predicted load trend and the actual trend. For example, by comparing whether the direction and slope of the predicted trend segment and the actual trend segment are consistent to calculate the accuracy rate.

[0100] The scene adaptability index is used to evaluate the performance of the prediction model under different actual scenarios. Under different data missing rate scenarios, the model is tested to see if it can maintain good prediction performance through data completion and feature extraction when there is different proportion of missing data (such as 10%, 20%, 30% missing rate) in the charging data. Under different weather scenarios, such as high temperature, low temperature, rain, and wind, the model's prediction accuracy is evaluated, for example, the change in MAE, MAPE, and other error indicators of the model under sunny and rainy conditions.

[0101] The robustness index focuses on the model's stable prediction ability when it is disturbed. In terms of data noise interference, different intensities of random noise (such as Gaussian noise, with noise intensity measured by signal-to-noise ratio) are added to the charging data to observe the fluctuation of the model's prediction results. In the hardware failure simulation scenario, such as simulating the failure of the charging pile data acquisition module, causing some period data to be abnormal or missing, the model's ability to effectively identify and exclude interference and maintain prediction stability is tested.

[0102] The economic index measures the economic impact on the power grid and users after developing demand response strategies based on the prediction results from the actual application value perspective. For changes in grid dispatching costs, the changes in grid peak shaving costs and power transmission loss costs after adjusting power generation plans, grid flow distribution, and other operations based on prediction results are analyzed. For changes in user charging costs, the savings or increases in user charging costs after adjusting charging times according to the predicted charging load peak and valley periods under time-of-use pricing and other policies are calculated.

[0103] After constructing these indexes, the weights of each evaluation index are determined by the Analytic Hierarchy Process (AHP). First, the evaluation target is divided into error, scene adaptability, robustness, and economicity as the criterion layer, and further divided into specific indexes as the scheme layer to construct a hierarchical structure model. Then, the judgment matrix is constructed by expert scoring method, the maximum eigenvalue and its corresponding eigenvector of the judgment matrix are calculated, consistency check is performed to ensure that the judgment matrix meets the logic, the weights of each evaluation index are obtained, and a complete weight system is formed according to the weights of each evaluation index; Next, for the target electric vehicle charging load prediction results, the actual scores under each evaluation index are calculated; Finally, the actual scores under each evaluation index are multiplied by the corresponding total weight, and the sum of all products is calculated to obtain the comprehensive evaluation score, which completes the weight integration and forms a comprehensive evaluation model that can directly input the prediction results and output the comprehensive evaluation results.

[0104] Then, the target electric vehicle charging load prediction result is input into the comprehensive evaluation model. The comprehensive evaluation model will weight the scores of the error index, the scene adaptability index, the robustness index and the economic index according to the weights of the indexes, and obtain a comprehensive evaluation result. If the comprehensive evaluation result reaches a preset standard, it indicates that the charging load prediction result has high reliability and application value, and at this time, a demand response strategy is formulated according to the target electric vehicle charging load prediction result, such as guiding users to staggered charging during the peak charging load prediction period, or adjusting the power supply plan of the power grid to ensure stable and economical power supply.

[0105] In some embodiments, a warning level is constructed according to the target electric vehicle charging load prediction result and the maximum bearable load threshold of the power grid, and the division standard of the warning level is dynamically adjusted according to the proportion of the predicted load exceeding the maximum bearable load threshold of the power grid and the load difference value change rate in the target electric vehicle charging load prediction result; Firstly, the maximum bearable load threshold of the power grid is determined, which is determined by the transmission capacity of the power grid, the capacity of the power transformation equipment, the line loss and other factors, and is a key index to ensure the safe and stable operation of the power grid. Then, a warning level is constructed according to the target electric vehicle charging load prediction result and the threshold.

[0106] Specifically, different warning intervals are divided according to the proportion of the predicted load exceeding the maximum bearable load threshold of the power grid, such as yellow warning when the exceeding proportion is 0-10%, orange warning when the exceeding proportion is 10-20%, and red warning when the exceeding proportion is more than 20%. At the same time, the division standard of the warning level is dynamically adjusted in combination with the load difference value change rate. The load difference value change rate reflects the change speed of the difference between the predicted load and the maximum bearable load threshold of the power grid. If the change rate is fast, it indicates that the load growth or decline trend is obvious, and the exceeding proportion range corresponding to each warning level needs to be adjusted appropriately to respond to the dynamic change of the power grid load more timely.

[0107] The user is stratified in combination with the user portrait and the user response behavior prediction model to obtain a user level; wherein the user portrait includes user charging frequency, charging time preference, charging power demand, price sensitivity and historical response record, and the user response behavior prediction model predicts the response probability of users at different levels under different incentive strategies based on historical data; specifically as follows: When user stratification is performed, a comprehensive user profile is first constructed. The user profile covers the user's charging frequency, i.e. the number of times of charging per unit time, which can reflect the user's use intensity of the electric vehicle; the charging time period preference, which understands which time period in a day the user habitually charges, which is crucial for peak-shaving charging guidance; the charging power demand, which is different for different vehicle models and different use scenarios, and the large power demand has a greater impact on the power grid load; the price sensitivity, i.e. the user's response degree to price changes, the price-sensitive user is more likely to be affected by the incentive strategy to adjust the charging behavior; and the historical response record, which records the user's participation and response degree in the past similar demand response scenarios.

[0108] Then, a user response behavior prediction model is built. The model is based on a large amount of historical data, including the response of each user under different incentive strategies (such as price discount, point reward, etc.), and is trained by machine learning algorithms (such as logistic regression, decision tree, etc.). The model can predict the response probability of each level of user (different levels according to the user profile) under different incentive strategies, providing data support for subsequent development of incentive strategies.

[0109] According to the user response probability obtained from the user response behavior prediction model, the incentive intensity is dynamically adjusted. For users with high response probability, the incentive intensity can be appropriately reduced, because even with small incentives, such users are more likely to respond; while for users with low response probability, the incentive intensity needs to be increased to improve the response willingness of such users. Combined with the early warning level and user level constructed in the foregoing, a differentiated incentive strategy is developed. For example, in the red early warning and high user charging power demand and low price sensitivity level, high price discount and point reward are provided to encourage the adjustment of charging time period; in the yellow early warning and low user charging frequency and flexible charging time period preference level, moderate point reward can be given to guide the peak-shaving charging.

[0110] According to the differentiated incentive strategy, an individual demand response scheme is obtained; based on the developed differentiated incentive strategy, an individual demand response scheme is generated for the specific situation of each user. The scheme informs the user in detail about what time period to adjust the charging behavior, how much price discount or point reward can be obtained, and the benefits to the power grid and himself after adjusting the charging behavior, etc., to ensure that the user can clearly understand and make a response decision.

[0111] An edge computing-based vehicle-network interaction platform is built. Edge computing has the characteristics of low delay and high bandwidth, and can perform data processing and communication at the network edge close to the electric vehicle terminal, the charging pile control system and the user APP. A low-delay communication protocol such as the Message Queuing Telemetry Transport (MQTT) protocol is used to realize real-time communication with the electric vehicle terminal, the charging pile control system and the user APP. Through the platform, the generated personalized demand response scheme can be pushed to the user in time to ensure the timeliness of information transmission and enable the user to quickly obtain and adjust the charging behavior according to the scheme.

[0112] A real-time feedback and dynamic adjustment mechanism is established. On the one hand, the response of the user to the personalized demand response scheme is collected, including whether to adjust the charging behavior according to the scheme, how to adjust the degree, etc.; on the other hand, the actual load data of the power grid is monitored in real time, and the difference between the predicted load and the actual load is compared. According to these feedback information, the range of responding users is dynamically adjusted, such as expanding or reducing the incentive coverage range of a certain type of users if the response effect of the users is not good; the differentiated incentive strategy is adjusted, such as modifying the discount ratio of electricity price, the rules of integral rewards, etc.; and the warning level is adjusted to ensure that the warning can accurately reflect the current load condition of the power grid. If the load reduction amount does not reach the pre-set target value, it means that the demand response effect is not as expected, and the demand response strategy needs to be optimized, such as increasing the incentive intensity, adjusting the incentive method, etc.; if the user dissatisfaction value exceeds the pre-set value, it means that the current strategy may bring inconvenience or dissatisfaction to the user, and the strategy also needs to be optimized to balance the demand response effect and user satisfaction while ensuring the load demand of the power grid.

[0113] In order to make the present application clearer, the present application provides a schematic diagram of the prediction result of the electric vehicle charging load, as shown in Figure 2 The red curve in the figure is the prediction value of the first model, and the blue curve is the actual true value. From the figure, it can be seen that within the time span of 0-100 15-minute periods, the prediction curve of the first model has a high overall fit with the true curve, especially in the key nodes of load change, such as the stage of rapid load rise around 20-40 periods, and the subsequent peak-valley change process, the prediction value can well follow the trend of the true value, the mean absolute percentage error (MAPE) is 4.7905%, which reflects that the model has good prediction ability for short-term electric vehicle charging load.

[0114] The present application provides a schematic diagram of the prediction error result, as shown in Figure 3As shown in the figure, the red curve in the figure represents the prediction error (the difference between the predicted value and the true value) of the Informer model. In the time span of 0-100 15-minute periods, the error fluctuates within a certain range, the absolute value of the error is relatively small in most periods, but there are large error peaks in some periods (such as the 40th period). However, overall, combined with the MAPE index of 4.7905%, it shows that the prediction error of the model is within an acceptable range, and at the same time, it also reflects that under some special load change scenarios, the model still has room for further optimization of the error.

[0115] The application also provides a comparison table of short-term electric vehicle charging load prediction results of the first model, the CNN model and the LSTM model, as shown in Table 1: Table 1: Comparison table of short-term electric vehicle charging load prediction results of the first model, the CNN model and the LSTM model

[0116] Based on the above, the application embodiment provides a low-data-quality-scenario electric vehicle charging load prediction method. In this method, first, the first feature affecting the charging load is obtained; a dynamic spatio-temporal weight factor is set according to the first feature affecting the charging load; a spatio-temporally correlated charging power matrix is constructed; the missing values in the spatio-temporally correlated charging power matrix are completed according to the dynamic spatio-temporal weight factor, to obtain complete target charging power matrix data; a three-dimensional input tensor is constructed according to the complete target charging power matrix data; the three-dimensional input tensor is input into the first model to obtain a charging load prediction value; the charging load prediction value is trend checked with historical charging load data to obtain a prediction trend; if the difference between the prediction trend and the historical trend exceeds a preset threshold, the charging load prediction value is linearly corrected based on the change rate of the historical trend to obtain a target electric vehicle charging load prediction result. By obtaining the first feature affecting the charging load and setting the dynamic spatio-temporal weight factor, the application can more accurately capture the spatio-temporal variation law of the charging load, thereby completing the missing values of the charging power matrix and constructing a three-dimensional input tensor that is more in line with reality. In combination with the first model prediction and the checking and correction of the prediction trend and the historical trend, the accuracy and reliability of the charging load prediction result are guaranteed from multiple dimensions, providing strong data support for efficient demand response, optimal resource allocation and the like of the power grid.

[0117] The above is combined Figures 1 to 3 The low-data-quality-scenario electric vehicle charging load prediction method provided by the application embodiment is described in detail, and the device and equipment provided by the application embodiment will be described below with reference to the accompanying drawings.

[0118] The application embodiment also provides a low-data-quality-scenario electric vehicle charging load prediction device, as shown in Figure 4As shown, the figure is a schematic diagram of a device for electric vehicle charging load prediction in a low data quality scenario according to an embodiment of the application. The device comprises: The acquisition module 401 is configured to acquire first features affecting the charging load. The completion module 402 is configured to set a dynamic spatio-temporal weight factor according to the first features affecting the charging load, construct a spatio-temporal correlated charging power matrix, complete the missing values in the spatio-temporal correlated charging power matrix according to the dynamic spatio-temporal weight factor, and obtain complete target charging power matrix data. The prediction module 403 is configured to construct a three-dimensional input tensor according to the complete target charging power matrix data, input the three-dimensional input tensor into a first model to obtain a charging load prediction value, perform trend checking on the charging load prediction value and historical charging load data to obtain a prediction trend, and perform linear correction on the charging load prediction value based on the change rate of the historical trend if the difference between the prediction trend and the historical trend exceeds a preset threshold to obtain a target electric vehicle charging load prediction result.

[0119] In some possible implementation manners, the acquisition module 401 is specifically configured to acquire first charging data from a plurality of electric vehicle charging stations and charging piles, acquire multi-source correlated influence data, identify abnormal values in the first charging data and the multi-source correlated influence data by using an abnormal value detection algorithm, eliminate the abnormal values, perform encoding conversion on text type correlated influence data in the first charging data and the multi-source correlated influence data after the abnormal values are eliminated, and unify the numerical value range to construct a first numerical feature set, construct a second numerical feature set according to non-text type correlated influence data in the first charging data and the multi-source correlated influence data after the abnormal values are eliminated, and filter the first numerical feature set and the second numerical feature set by using feature correlation analysis to obtain the first features affecting the charging load.

[0120] In some possible implementation manners, the completion module 402 is specifically configured to construct a low-rank matrix completion optimization model based on the dynamic spatio-temporal weight factor, solve the low-rank matrix completion optimization model by using a singular value threshold algorithm to obtain a target value, replace the missing values in the spatio-temporally correlated charging power matrix with the target value to obtain a completed charging power matrix, and perform checking on the completed charging power matrix to obtain a result of passing the checking, and take the completed charging power matrix as the target charging power matrix data.

[0121] In some possible implementation manners, the device further comprises: The optimization module is configured to construct an initial neural network model, and optimize the initial neural network model by using a pollination algorithm to obtain the first model.

[0122] In some possible implementation manners, the device further comprises: The evaluation module is configured to construct a full-scenario multi-dimensional evaluation system including error indicators, scenario adaptability indicators, robustness indicators and economic indicators; determine the weights of the evaluation indicators by using the analytic hierarchy process; obtain actual scores under each evaluation layer indicator according to the target electric vehicle charging load prediction result; obtain a comprehensive evaluation result according to the weights of the evaluation indicators and the actual scores under each evaluation layer indicator; and if the comprehensive evaluation result reaches a preset standard, formulate a demand response strategy according to the target electric vehicle charging load prediction result.

[0123] In some possible implementation manners, the evaluation module is specifically configured to construct an early warning level according to the target electric vehicle charging load prediction result and the maximum bearable load threshold of the power grid, dynamically adjust the division standard of the early warning level according to a proportion of a predicted load exceeding the maximum bearable load threshold of the power grid in the target electric vehicle charging load prediction result and a load difference change rate, perform user stratification in combination with a user portrait and a user response behavior prediction model to obtain a user level, wherein the user portrait includes a user charging frequency, a charging time period preference, a charging power demand, a price sensitivity degree and a historical response record, and the user response behavior prediction model predicts a response probability of each level user under different incentive strategies based on historical data; dynamically adjust an incentive intensity according to the response probability of the user, and formulate a demand response strategy according to the incentive intensity, the early warning level and the user level.

[0124] In some possible implementation manners, the dynamic spatiotemporal weight factor includes a region weight and a time period weight, wherein the region weight is dynamically adjusted according to a historical load contribution degree and an operation state of a corresponding region charging station in the first feature of the influence on the charging load, and the time period weight is dynamically configured according to a charging load fluctuation feature in different time periods and a real-time electricity demand change in the first feature of the influence on the charging load.

[0125] The electric vehicle charging load prediction device in the low data quality scenario according to the embodiments of the present application can correspond to the method described in the embodiments of the present application, and the above other operations and / or functions of each module / unit of the electric vehicle charging load prediction device in the low data quality scenario are respectively implemented to realize Figure 1 The above other operations and / or functions of each module / unit of the electric vehicle charging load prediction device in the low data quality scenario are respectively implemented to realize

[0126] The embodiments of the present application also provide a computing device. As Figure 5 shown, the figure is a schematic diagram of a computing device according to the embodiments of the present application. The computing device 500 includes a bus 501, a processor 502, a communication interface 503 and a memory 504. The processor 502, the memory 504 and the communication interface 503 communicate through the bus 501.

[0127] Bus 501 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0128] The processor 502 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0129] Communication interface 503 is used for communication with external devices.

[0130] Memory 504 may include volatile memory, such as random access memory (RAM). Memory 504 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0131] The memory 504 stores executable code, and the processor 502 executes the executable code to perform the aforementioned electric vehicle charging load prediction method under low data quality scenarios.

[0132] Specifically, in achieving Figure 4 In the case of the illustrated embodiment, and Figure 4 When the modules or units of the electric vehicle charging load prediction device in the low data quality scenario described in the embodiment are implemented by software, the following steps are performed: Figure 4 The software or program code required for the functions of each module / unit can be partially or entirely stored in memory 504. Processor 502 executes the program code corresponding to each unit stored in memory 504 to execute the aforementioned electric vehicle charging load prediction method under low data quality scenarios.

[0133] The embodiment of the present application further provides a computer readable storage medium. The computer readable storage medium can be any available medium or data storage device that can be used to store instructions that can be executed by a computing device. The computer readable storage medium can be a magnetic medium (e.g., a floppy diskette, a hard disk drive), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid state hard drive) or any other medium that can be used to store instructions that can be executed by a computing device. The computer readable storage medium includes instructions that instruct a computing device to perform the method for predicting charging load of an electric vehicle in a low data quality scenario.

[0134] The embodiment of the present application further provides a computer program product including one or more computer instructions. When the computer instructions are loaded and executed on a computing device, the processes or functions described in the embodiment of the present application are generated in whole or in part.

[0135] The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer or data center to another website, computer or data center through a wired (e.g., coaxial cable, optical fiber) or wireless (e.g., infrared, wireless, microwave, etc.) manner.

[0136] The computer program product is executed by a computer, and the computer executes any of the methods for predicting charging load of an electric vehicle in a low data quality scenario. The computer program product can be a software installation package, and when any of the methods for predicting charging load of an electric vehicle in a low data quality scenario is needed, the computer program product can be downloaded and executed on the computer.

[0137] The description of the processes or structures corresponding to the above respective figures has different focuses, and the parts not described in detail in a certain process or structure can be referred to the related description of other processes or structures.

[0138] The above is merely a specific implementation of the present application, but the protection scope of the present application is not limited to this. Any change or replacement within the technical scope disclosed in the present application should be covered in the protection scope of the present application.

Claims

1. A method for electric vehicle charging load forecasting under low data quality, characterized in that, The method comprises: obtaining a first feature affecting the charging load; setting a dynamic space-time weight factor according to the first feature affecting the charging load; constructing a space-time correlated charging power matrix; completing the missing values in the space-time correlated charging power matrix according to the dynamic space-time weight factor to obtain complete target charging power matrix data; constructing a three-dimensional input tensor according to the complete target charging power matrix data; inputting the three-dimensional input tensor into a first model to obtain a charging load prediction value; trend checking the charging load prediction value with historical charging load data to obtain a prediction trend; if the difference between the prediction trend and the historical trend exceeds a preset threshold, linearly correcting the charging load prediction value based on the change rate of the historical trend to obtain a target electric vehicle charging load prediction result.

2. The method of claim 1, wherein, The method comprises: obtaining a first feature affecting the charging load; obtaining first charging data from a plurality of electric vehicle charging stations and charging piles; obtaining multi-source correlated influence data; using an outlier detection algorithm to identify outliers in the first charging data and the multi-source correlated influence data, and removing the outliers; encoding and converting the text type correlated influence data in the first charging data and the multi-source correlated influence data after removing the outliers, and unifying the numerical value range to construct a first numerical feature set; constructing a second numerical feature set from the non-text type correlated influence data in the first charging data and the multi-source correlated influence data after removing the outliers; 3. The method of claim 1, wherein, screening the first numerical feature set and the second numerical feature set through feature correlation analysis to obtain the first feature affecting the charging load. The method comprises: constructing a low-rank matrix completion optimization model based on the dynamic space-time weight factor; solving the low-rank matrix completion optimization model using a singular value threshold algorithm to obtain a target value, replacing the missing values in the space-time correlated charging power matrix with the target value to obtain a completed charging power matrix; 4. The method of claim 1, wherein, verifying the completed charging power matrix to obtain a verification passed result, and taking the completed charging power matrix as the target charging power matrix data. The method further comprises: constructing an initial neural network model; 5. The method of claim 1, wherein, optimizing the initial neural network model using a pollination algorithm to obtain a first model. The method further comprises: constructing a full-scenario multi-dimensional evaluation system including error indicators, scene adaptability indicators, robustness indicators, and economic indicators; determining the weights of each evaluation indicator through an analytic hierarchy process; obtaining actual scores under each evaluation layer indicator according to the target electric vehicle charging load prediction result; obtaining a comprehensive evaluation result according to the weights of each evaluation indicator and the actual scores under each evaluation layer indicator; 6. The method of claim 5, wherein, if the comprehensive evaluation result meets a preset standard, formulating a demand response strategy according to the target electric vehicle charging load prediction result. The method comprises: An early warning level is constructed according to the target electric vehicle charging load prediction result and a maximum bearable load threshold of a power grid, and a division standard of the early warning level is dynamically adjusted according to a proportion of a predicted load exceeding the maximum bearable load threshold of the power grid and a load difference rate in the target electric vehicle charging load prediction result; User stratification is performed in combination with a user portrait and a user response behavior prediction model to obtain a user level; the user portrait includes a user charging frequency, a charging time period preference, a charging power demand, a price sensitivity degree, and a historical response record, and the user response behavior prediction model predicts a response probability of each level user under different incentive strategies based on historical data; An incentive intensity is dynamically adjusted according to the response probability of the user, and a demand response strategy is formulated according to the incentive intensity, the early warning level, and the user level.

7. The method of claim 1, wherein, The dynamic spatiotemporal weight factor includes a region weight and a time period weight; the region weight is dynamically adjusted according to a historical load contribution degree and an operation state of a corresponding region charging station in the first feature of the charging load influence, and the time period weight is dynamically configured according to a charging load fluctuation feature and a real-time electricity demand change in different time periods in the first feature of the charging load influence.

8. A low-data quality scenario-based electric vehicle charging load prediction device, characterized by, The device includes: An acquisition module configured to acquire a first feature of charging load influence; A completion module configured to set a dynamic spatiotemporal weight factor according to the first feature of charging load influence, construct a spatiotemporally associated charging power matrix, complete missing values in the spatiotemporally associated charging power matrix according to the dynamic spatiotemporal weight factor, and obtain complete target charging power matrix data; A prediction module configured to construct a three-dimensional input tensor according to the complete target charging power matrix data, input the three-dimensional input tensor into a first model to obtain a charging load prediction value, perform trend verification on the charging load prediction value and historical charging load data to obtain a prediction trend, and perform linear correction on the charging load prediction value based on a change rate of a historical trend if a difference between the prediction trend and the historical trend exceeds a preset threshold to obtain a target electric vehicle charging load prediction result.

9. A computing device, comprising: includes a memory and a processor; One or more computer programs are stored in the memory, and the one or more computer programs include instructions; when the instructions are executed by the processor, the computing device performs the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium is used to store a computer program for executing the method of any one of claims 1 to 7.

Images